Feature Engineering Approach to Detect Retinal Vein Occlusion Using Ultra-Wide Field Fundus Images

Retinal Vein Occlusion (RVO) symptoms can be identified through analysis of fundus image capturing the retinal area of an eye. One of the symptoms associated with RVO is haemorrhage lesions due to rapture or damage of blood vessels in the retina. This paper investigates the feature values associated with haemorrhage symptoms in wide-angle fundus images. The feature values are used to construct a classifier to label an image into RVO or non-RVO. A total of 80 feature values are extracted based on various image properties. Four classifiers are built by using the feature vectors to compare their performance. A total of 87 wide-angle images are used in the evaluation. It is found that shape- and colour-based features are useful for separating the RVO images from non-RVO images with a sensitivity of 0.80 and specificity of 0.85. Given the accuracy rate of 0.90, specificity of 0.92, and sensitivity of 0.88 on RVO detection, SVM performed best compared to other classifiers. The traditional feature-based approach can achieve performance levels close to the deep learning approaches using UWF images for haemorrhage prediction.